I contribute to the tools I use. That includes merged fixes in Apple's MLX and MLX-LM, along with ongoing work on Unsloth Studio, MTPLX, and oMLX.

I've also submitted PRs to MLX Serve, vLLM Metal, dflash, Google Ads MCP, Nixpkgs, conda-forge, MacPorts, Rust, and Bootstrap.

See my upstream pull requests

I trained Wisp Coder from tokenizer to released weights on Apple Silicon: a 32K-token tokenizer, 108.2M parameters, and 5 billion training tokens, with fill-in-the-middle and native multi-token prediction. The trunk has 100.7M parameters; 108.2M includes the MTP module.

Try this: open the demo, edit the code around the cursor, and ask Wisp to fill the gap. The free CPU demo uses the standard decoder.

Model card & weights ↗ · Training case study ↗ · Portfolio story ↗

The record includes comparisons that didn’t favor Wisp. The packaged MTP runtime checks correctness; the release does not claim a production MTP speedup.

MCP integrations, reusable skills, identity, and approval workflows for the systems a business already uses. At Basilecom / Transmission Agency, my platform work supports 120 staff across the UK, US, and APJ, with 12+ production MCP integrations.

The current open-source direction brings that work onto the desktop: isolated coding workspaces, bounded execution, and one visible browser a person and an agent can use together. The shared-browser work is in progress, with recovery and human takeover part of what is being built and qualified.

Explore this: follow how access, tool contracts, operational controls, and adoption fit together in the case study.

Engineering case study ↗ · MCP is a governance problem ↗

MLX conversions, native multi-token prediction in MTPLX, and SSD-backed expert streaming with iliria. The work spans runtime behavior, weights, memory, and storage.

Explore this: start with the architecture, then inspect source revisions, runtime recipes, and published measurement records before choosing a model.

iliria source ↗ · Measuring inference fairly ↗ · Apple Silicon AI catalog ↗

RAG systems with evaluation across retrieval, answer grounding, latency, cost, and regression behavior. The useful part is the release process around the answer.

Explore this: follow the retrieval and evaluation loop, including the checks used to catch regressions before release.

Learn by building the loop ↗ · More engineering case studies ↗

• iliria: C/Metal inference that streams large MoE models from SSD, built on colibri. I also published the GLM-5.2 int4 container it serves.
• racecontrol: routing and failure recovery across local inference engines, with a runnable HTTP demo.
• CallSieve: local code retrieval for coding agents, with CLI and MCP interfaces.
• VecStore: embedded vector search with metadata filtering and persistence.
• PhilJS: an experimental TypeScript UI framework with a dependency-free signals demo.

More projects, model experiments, and creative work

Browse the engineering case studies.

Five examples of the work behind the project list:

• Fixing a Metal quantization bug in Apple MLX: the boundary case, root cause, final upstream change, and regression tests.
• Training and releasing Wisp Coder: tokenizer training, model pretraining, runtime compatibility, and controlled experiments.
• Taking training data into model releases: data imports with source records, verification tools, expert-pruning calibration, and measured model tradeoffs.
• Measuring local inference fairly: matching outputs and workloads, separating streaming behavior from speed, and reporting the actual margin.
• Connecting enterprise systems to AI agents: integrations, identity, operational controls, and adoption across a global agency.

My Hugging Face work covers original models, conversions, experimental derivatives, and training data. I work on both the models themselves and the engineering needed to run them locally.

20 public model repositories · 1 public dataset · 2 public Spaces · 11,388 monthly model downloads

Hugging Face Hub snapshot, September 11, 2026; downloads summed across my public model repositories.

Explore my Local AI Guide. I built a searchable catalog of the releases so you can browse by project, use case, artifact type, and download size. Each entry links to the model card, the inspected revision, and available evaluation records, with the runtime requirements explained alongside it.

Try Wisp in your browser. Edit the code around the cursor and ask it to fill the gap. The free demo runs the standard decoder on CPU.

With Wisp Coder 110M, I took the work from tokenizer training to released weights. I trained a 32K-token tokenizer on 400,000 documents and a 108.2M-parameter code model on 5 billion tokens using MLX on Apple Silicon. Fill-in-the-middle and multi-token prediction were part of training from the start.

I checked the export against Transformers, published decoding correctness checks, and evaluated five models across 1,372 code-completion tasks. I published the comparisons even when they didn't favor Wisp. The Wisp case study covers the design decisions, evidence, and limitations.

• Hy3 and GLM-5.2 compression — Pruned mixture-of-experts models, trained LoRA adapters, and published smaller MLX builds. The Hy3 release removes 25% of experts per layer, then uses LoRA training to recover capability. The cards include evaluations, regressions, and the exact runtime recipes.
• Qwen Fable-Fusion for MLX and MTPLX — Reconstructed DavidAU's GGUF release for MLX, published 4-, 6-, and 8-bit builds, and calibrated the multi-token prediction head. The featured 6-bit build includes vision support through MTPLX, with conversion details and measured decoding results in the card.
• Ornith 1.5 for MTPLX — Converted Ornith AI's model for Apple Silicon with mixed-precision weights and a BF16 MTP head. The card documents source provenance, runtime requirements, and what has and hasn't been validated.
• Akka for MLX — Converted nightmedia's merge to 6-bit MLX and checked its draft head against the target model. Calibration failed, so I released it without the MTP head and published the results.
• DeepSeek V4 Flash for MLX — Built an experimental conversion and published comparisons against the original model, including a full 198-question GPQA Diamond run. Retained as a reference, with the quality regressions and faster alternative documented.

I published GLM-5.2 Demolition training and calibration data: 154 JSONL files covering code training, agent tool-use examples, repair, domain-specific adapters, and expert-pruning calibration.

I also built the import and verification tools around it: sampling by domain, normalizing chat and tool messages, keeping source labels, and separating training examples from calibration prompts. The data-to-model case study follows the public code and records into the Hy3 model experiments.

Explore the datasets, training methods, and all 20 model repositories

I also group featured releases in Selected Work and research artifacts in Apple Silicon — Experimental Models.

Browse all models · Browse datasets

I started building university websites at Fordham. Since then, I've worked on IBM's enterprise search, Atlas Air's flight scheduling, Dragos's cybersecurity products, and U.S. Air Force mission-planning software through client engagements. My healthcare work includes Teladoc and IntegraMed; earlier, I helped build commerce at BaubleBar and campaign experiences at 360i.

That range matters when a project gets complicated. I've had to balance user experience, sensitive data, uptime, budgets, and deadlines—and help other engineers do the same.

At Teladoc, I worked on telehealth through its NYSE debut. At IntegraMed, on clinical systems and machine learning. The environments changed, but the work kept coming back to user experience, sensitive data, performance, and dependable releases.

A few examples from my recent work:

• AI platform ownership: Architecture and technical direction for a global agency, supporting 120 staff across the UK, US, and APJ, including 12+ production MCP integrations.
• Operating cost: A Snowflake permissions and governance cleanup that reduced credit consumption by 35%, about £2,800 a month.
• Team development: Led teams of 4–20 and coached five engineers into senior roles.

I’m based in New Rochelle. Away from software, I’ve volunteered with the Civil Air Patrol, photographed staff portraits pro bono for Pelham Children’s Center, and spent years involved in youth hockey.

Full career record · Career background and case studies

A useful research record includes the misses. Wisp’s published comparisons include results that favored other models. The Akka conversion shipped without its MTP head after calibration failed. The experimental DeepSeek conversion documents quality regressions and a faster alternative.

That’s part of leaving a trail someone else can actually use.